scVI hyperparameter tuning

[mlin]

Our regular procedure for generating Census scVI embeddings trains the scVI model with an up-to-date HVG h5ad, but doesn't revisit the hyperparameters set a couple of years ago, since which Census has grown a lot (and scVI itself has likely evolved).

This notebook is a record of the "autotune" done in 2023: https://github.com/YosefLab/census-scvi/blob/main/experiments/autotune/notebooks/2023_09_autotune_report.ipynb

And some historical discussion: #777

But to repeat, we'd still need to recover:

• detailed recipe for generating the autotune results analyzed there (installation, commands, etc.)
• expertise/judgment to choose the best parameters from the results